pascalandy/transcript-sk

Transcribe YouTube videos via Deepgram with optional OpenCode, Codex, or Claude summary.

Requires Deepgram API key in keyring:

chezmoi secret keyring set --service=deepgram --user=api_key

Run

uv run ~/.config/opencode/skills/transcript-sk/scripts/transcript.py "<youtube_url>"

Use a 10-minute timeout (600 seconds) when running this script.

Discovery (no URL required):

uv run ~/.config/opencode/skills/transcript-sk/scripts/transcript.py --list-prompts
uv run ~/.config/opencode/skills/transcript-sk/scripts/transcript.py --list-models
uv run ~/.config/opencode/skills/transcript-sk/scripts/transcript.py --help

Choose flags from user intent

Default behavior: Script defaults to follow_along_note prompt with opencode agent 1-kimi. Just provide the URL.

• Default (user gives URL only) → run with just <url> (script defaults to opencode + 1-kimi + follow_along_note)
• Use OpenCode with a different agent → add --provider opencode --model <agent> (default: 1-kimi)
• Use Codex → add --provider codex
• Transcript only (user explicitly says "transcript only" or "no summary") → add --no-prompt
• Different prompt requested (user says short_summary, summary_with_quotes, etc.) → add --prompt <stem>
• Codex model requested → add --provider codex --model <name> (default: gpt-5.4, reasoning: high)
• Claude model requested (user says haiku/sonnet/opus) → add:
  • haiku → --model claude-haiku-4-5
  • sonnet → --model claude-sonnet-4-5
  • opus → --model claude-opus-4-5
• If user asks "what prompts/templates exist?" → run --list-prompts, return the list, and ask which one to use.
• If user asks "what models/agents can I use?" → run --list-models (and set --provider claude if they ask for Claude models), return the list, and ask which one to use.

Shorthand input mapping

Examples:

# User: transcript https://www.youtube.com/watch?v=2QpXab8z_Gw
# (default: opencode + 1-kimi + follow_along_note prompt)
uv run ~/.config/opencode/skills/transcript-sk/scripts/transcript.py "https://www.youtube.com/watch?v=2QpXab8z_Gw"

# User: transcript, opencode, gpthigh, short_summary, https://www.youtube.com/watch?v=2QpXab8z_Gw
uv run ~/.config/opencode/skills/transcript-sk/scripts/transcript.py --provider opencode --prompt short_summary --model gpthigh "https://www.youtube.com/watch?v=2QpXab8z_Gw"

# User: transcript, short_summary, claude, haiku, https://www.youtube.com/watch?v=2QpXab8z_Gw
uv run ~/.config/opencode/skills/transcript-sk/scripts/transcript.py --provider claude --prompt short_summary --model claude-haiku-4-5 "https://www.youtube.com/watch?v=2QpXab8z_Gw"

# User: transcript only, https://www.youtube.com/watch?v=2QpXab8z_Gw
uv run ~/.config/opencode/skills/transcript-sk/scripts/transcript.py --no-prompt "https://www.youtube.com/watch?v=2QpXab8z_Gw"

Outputs

• The script creates a timestamped output folder (it prints the path and opens it in Finder on macOS).
• Key files: {prompt}.md, raw_transcript.txt, raw_sentences.txt, raw_transcript.json, meta.txt.

For runtime requirements and validation steps, see ~/.config/opencode/skills/transcript-sk/README.md.

Don't READ the final answer. The user can already see it in the terminal.